Restoring the two pivotal fluoxetine trials in children and adolescents with depression

3.1.1. Randomisation and blinding

Randomisation was stratified by gender and age [9]. There was no information on block size. Treatment assignment was carried out by a local pharmacist using a list prepared by the biostatistician. A study site nurse verified the assignment was correct by comparing dispensed medication to the list. This suggests the allocation was not concealed.

The pharmacy initially blinded the drugs by emptying Prozac capsules for the placebo group and refilling them with lactose powder [9]. From August 1993, Lilly supplied blinded drugs in white capsules.

The hospital pharmacy informed the chemistry labs about treatment assignment “to avoid running unnecessary blood levels [9]”. A study site nurse who checked the study medication liaised between the clinical site and the pharmacy and provided psychiatric ratings for two patients. Two patients who attempted suicide on fluoxetine “may have had their treatment assignments revealed,” and there were “a few instances … (less than ten)” where study records indicated physicians were unblinded prior to study completion. For patients who dropped out or became depressed, the treating clinician had access to their treatment assignment [9].

3.1.2. Outcomes

There were two primary psychiatrist rated outcomes: improvement on the Clinical Global Impressions (CGI) scale and on the CDRS-R [7]. These were reduced in the published article to binary outcomes, very much or much improved on the CGI scale, and a CDRS-R ≤ 28. The CSR had only one primary outcome, at least a 30% reduction from baseline on the CDRS-R, but there were three success criteria: remission (CDRS-R ≤ 28); response (CGI-Improvement score of 1 or 2); and recovery (both criteria) [9].

Secondary outcomes reported in the CSR or published article included:

Survival analysis of remission on the CGI scale [7]

Subgroup analyses included:

Analyses by gender, age and origin of trial drugs [9]

Adverse events were collected by asking the patients if they had any of 32 symptoms on a Side-Effects Checklist (described as 30 but there were 32), or any of 30 symptoms on the Fluoxetine Side-Effects Checklist (starting in January 1993), and by collecting non-solicited adverse events [9]. Adverse events were coded according to COSTART (Coding Symbols for a Thesaurus of Adverse Reaction Terms) by blinded Lilly personnel. Subgroup analyses of adverse events were performed for age, gender, and provider of trial drugs.

3.1.3. Washout phase

The patients were randomised after a one-week washout period on single-blind placebo [9]. As 5 fluoxetine versus 9 placebo patients were on tricyclic antidepressants [9], the risk of withdrawal symptoms was greatest for patients randomised to placebo.

3.1.4. Effect of fluoxetine on depression

The efficacy analyses favoured fluoxetine. After 4 weeks, 6 patients had discontinued on fluoxetine and 12 on placebo [9]. As most analyses used the last observation carried forward (LOCF) method, more patients on placebo than on fluoxetine had high depression scores carried forward.

After 8 weeks, the difference from baseline in CDRS-R was 9.7 larger on fluoxetine than on placebo using the LOCF method. It was 2.5 using observed cases on one graph and 6.0 on another graph on the same page with no explanation for this discrepancy.

The published article left out results for BID and FGAS. The CSR left out both of these results and those for WSAS and CGAS, stating the self-report scales were not collected consistently (but providing no evidence for this) and because “the data came from relatively unvalidated scales”.

Lilly did not explain discrepant results for BPRS-C in the CSR and the publication, e.g. on fluoxetine, the average score after treatment was 38.9 and 18.0 respectively, and the P-values were also markedly different, P = 0.32 (calculated by us) versus P = 0.09 (Table 1).

The published article combined the results for the CDI and BDI scales, but all averages were smaller than the weighted averages for the CDI and BDI scales in the CSR (see Table 1), which is a mathematical impossibility.

Despite the bias introduced by using exit values, there were no significant differences in BPRS-C (P = 0.32), CGAS (P = 0.18), or in outcomes reported by the patients, CDI/BDI (P = 0.58) and WSAS (P = 0.17) (Table 1) [7]. Emslie et al. stated that, “given the wide variability of initial child self-reports, these findings are difficult to interpret [7]”. However, the coefficients of variation (standard deviations divided by the means) were smaller for BPRS-C and CGAS than for CDRS-R (Table 1).

The CSR provided P-values for all individual 21 items on the BPRS-C and emphasized that fluoxetine increased hyperactivity significantly (P < 0.001) [9]. Hyperactivity is a warning signal for suicide and violence [2], but Lilly introduced the opposite concept: “A decrease of hypoactivity… which is associated with the patients returning to normal functioning at greater rates than placebo-treated patients” transforming a potentially harmful result into a benefit. There were no data in support of this interpretation and Emslie et al. stated: “whether long-term treatment would result in the amelioration of school, general functioning, or concurrent comorbidities is unknown [7]”.

No statistical adjustments can substitute missing data reliably. We therefore focussed on those patients with minimal symptoms after 8 weeks. Only 15 versus 11 patients had minimal symptoms (P = 0.49, our calculation), defined by Lilly as CDRS-R ≤ 28, and only 14 versus 9 had recovered (P = 0.34), defined as CDRS-R ≤ 28 and a CGI Improvement score of 1 or 2 [9]. Even these results may be biased in favour of fluoxetine, as spontaneous remission is common and more data were missing for patients on placebo than for patients on fluoxetine (23 versus 15 had dropped out by week 8).

The information on discontinuations and suicide attempts was inconsistent. In the published report [7], 36 patients discontinued, in the CSR 38 [9]. The numbers, reasons, and timing were the same in only 5 of 28 cells with information (Table 2). There was no explanation for these discrepancies [9].

No patients died. Two patients attempted suicide on fluoxetine, after 12 and 15 days, respectively [9]. These suicide attempts were left out from the published report [7]. One of them led to study discontinuation after 15 days, and it seems that Emslie et al. called this a “protocol violation” (see Table 2). Adverse events definitely or possibly predisposing to suicide in this patient according to our criteria were manic reaction, insomnia, and nervousness (possibly akathisia) [9]. Predisposing events for the other patient were anxiety, depression, neurosis, thinking abnormal, asthenia, and also hyperkinesia (features of akathisia).

Four additional patients discontinued fluoxetine because of adverse events called “minimal” in the published report, even though three of them developed manic symptoms and the fourth had a severe rash [7]. The CSR summary described two patients with hypomania, one with increased impulsivity, and one with rash [9]. The narratives on these patients were extremely brief, between 17 and 35 words [9]. One also had an anxiety attack and the other had hyperkinesia. The patient with increased impulsivity had hyperkinesia, anxiety and depression but was coded as personality disorder.

The CSR stated that “No placebo-treated patients discontinued due to adverse events” while the published article mentioned one such patient but not the event that caused discontinuation. In 2004, Emslie et al. published additional data, mentioning these two suicide attempts but ascribing one of them to the placebo group, referenced as “Emslie, personal communication [12]”. They also claimed that another patient on placebo discontinued because of mania even though this patient had received fluoxetine.

The CSR had more patients dropping out on placebo than on fluoxetine due to lack of effect (19 versus 6, P = 0.005) but fewer patients dropping out due to adverse events (0 versus 5, P = 0.056). Only one cause was assigned for each patient [7,9]. Adding both reasons, 19 versus 11 patients dropped out (P = 0.12, our calculation).

Concomitant drugs with sedative properties, often used to manage adverse events linked to suicide and violence, were mentioned for 11 patients on fluoxetine and for 3 patients on placebo.

One patient on fluoxetine and three on placebo did not experience any of the 32 adverse events listed in the Side-Effects Checklist [9].

The data on non-solicited adverse events and events on the Fluoxetine Side-Effects Checklist were not included in the main text of the CSR but came after “Discussion and Overall Conclusions” under the heading “Tables, Figures, and Graphics Not Included in Other Sections [9]”.

In the table of non-solicited adverse events, Lilly provided 29 P-values and wrote that there were no statistically significant differences but without noting that more patients on fluoxetine than on placebo experienced one or more events (P = 0.051, Lilly’s P-value) [9].

In contrast, a table showing 28 of the 30 items listed in the Fluoxetine Side-Effects Checklist did not present a single P-value even though there were conspicuous differences. For example, 32 fluoxetine versus 18 placebo patients experienced at least one adverse event (P = 0.008), 19 versus 6 experienced restlessness (P = 0.005), 9 versus 1 had nightmares (P = 0.02), and 7 versus 4 felt tense inside. Restless ness, including feeling tense inside, and nightmares increase the risk of suicide and violence [4,5]. Lilly argued that this checklist was not used consistently throughout the study and was not administered to all randomised patients but did not provide information on the number of missing values. There were no data on the severity of the adverse events even though it was assessed [9].

The published article did not present any data obtained in one of the three ways just described. Only 49 of the 1099 words (4%) in the Results section were related to safety and they were only about discontinued patients [7].

The published report stated electrocardiograms were taken at baseline and after 4 and 8 weeks, but according to the CSR only at baseline and “were not analyzed statistically”. The data were not reported in either the published article or the CSR.

3.2.1. Randomisation and blinding

Randomisation was stratified by gender and age using a computer-generated randomisation sequence. There was no information on block size. Sealed codes were available, which risks violating allocation concealment.

The medication was packaged in blister cards with a package number. All patients received three capsules daily containing 10 or 20 mg fluoxetine or matched placebos. Patients who did not tolerate fluoxetine could have their dose reduced from 20 to 10 mg or from 40 to 20 mg.

It is not clear if the blinding was maintained after the first 9 weeks when patients on placebo and responders on 20 mg fluoxetine continued with their drug while non-responders on fluoxetine were re-randomised to 20 or 40 mg fluoxetine.

“A minimal number of Lilly personnel” would see the randomisation table and codes before the study was complete. This minimum included statisticians, regulatory scientists, systems analysts, people working with pharmacokinetics, clinical laboratory medicine personnel, and medical writers, all of whom had unblinded access to data during the study. Patients dropping out could receive rescue therapy from a physician who was not blinded but “agreed to maintain the study blind from study personnel”. Lilly reported to the FDA that in their review of the source records it was not uncommon to see notations defining the patient’s blinded treatment, or in some cases to find fluoxetine plasma concentration results [13].

3.2.2. Outcomes

All study objectives centered on efficacy; none on safety. The primary outcome, CDRS-R, was dichotomized into at least a 30% reduction from baseline.

Numerous secondary objectives, outcomes and comparisons were described and what was reported was not always what was planned.

In the protocol, the secondary objectives “are as follows” whereas in the CSR, they “included the following,” suggesting there could have been more than those listed, which was indeed the case. CDRS-R remission rates and CGI response rates were exploratory “additional analyses” in the protocol and would be “conducted as deemed appropriate,” but in the CSR they were “also analyzed,” which term Lilly used for pre-planned analyses. Thus, protocol defined exploratory analyses acquired a status of secondary outcomes in the CSR.

Rating scales and analyses were changed. Mean score in the protocol became mean change for seven outcomes in the CSR and in another place in the protocol, and subgroup analyses by age became subgroup analyses by age, gender, and family history of depression.

Efficacy outcomes were:

Mean change in CDRS-R

Clinician rated outcomes were recorded at every single visit after randomisation (11 times) whereas the patient rated outcome (CDI or BDI, depending on age) was only recorded at 5 visits.

Non-solicited adverse events were collected by questioning the patients about the presence of adverse events “in a non-directed manner,” with no further instructions. A form for registering all adverse events that had occurred since the last visit and all clinically relevant abnormalities found on the physical exam or ECG took up one page. It had a 4.5 cm² box for each narrative, and three severity codes (mild, moderate, and severe) that were not defined anywhere. The forms for benefits took up 10 pages. There was an optional one-page form for comments on benefits or aberrant laboratory values, but not adverse events.

Solicited adverse events were captured with the Side-Effects Checklist. Verbatim terms were coded as COSTART terms by blinded Lilly personnel. The Fluoxetine Side-Effects Checklist developed by Lilly and used in trial X065 was not used.

Blood pressure, heart rate, height and weight were measured. The text noted that ECGs were taken at various times during the study. A schedule of events noted that they were taken only at the first visit and after 19 weeks, but a table of missing measurements showed that only 41 recordings had been missed, at various numbered visits throughout the whole trial.

3.2.3. Statistical analyses

A large number of analyses were carried out that were not prespecified in the protocol or in the statistical analysis plan, even though it was updated on 7 February 2000, two months after the 19 weeks trial was over, when a lot of Lilly staff had seen the unblinded values.

We found problematic uses of statistics and contradictory information. It was stated that all patients with at least one post-randomisation visit would be included in the efficacy analyses, but this did not apply to the primary efficacy analyses for which two visits were needed, including “at least 1 week taking fluoxetine 20 mg/day”.

Contrary to the company’s claim to have conducted intention-to-treat analyses, some patients with post-randomisation visits were excluded and the patients who were discontinued were not assessed at the end of the trial. As it was stated patients would stay in the groups to which they had been randomly assigned even if they did not follow the protocol, it made no sense to discontinue patients, the only effect of which was to bias the study results in favour of fluoxetine.

CSR analyses included:

10%, 20%, 30%, 40%, 50%, 60%, and 70% reductions in CDRS-R (only 30% was prespecified in the protocol)

Later in the CSR, there were additional analyses, e.g. for 4 CDRS-R subtotals and for each of the 17 individual items. Subgroup analyses were performed on CDRS-R response and mean change; CGI-Improvement; non-solicited and solicited adverse events, for children and adolescents separately, males and females, and for patients with and without a family history of depression. There were also results, with P-values, for each of 15 investigator sites for CDRS-R, CGI-Improvement, and MADRS.

Furthermore, “After reviewing some of the results of this study, additional exploratory analyses were performed. These included additional subgroup analyses and “completers” analysis of laboratory, ECG, and vital signs data… A second set of blinded ECG readings performed by a pediatric cardiologist supplemented the original readings”.

Even though only 34% of the patients completed all 19 weeks, the analyses after 19 weeks were called interim analyses. All the analyses performed after 9 weeks were repeated after 19 weeks along with additional ones, e.g. including those patients who received “at least 4 weeks of fluoxetine 20 mg/day”. As nothing was said about the placebo group, it is not clear if the same criteria applied to the placebo group, or if all placebo patients were included, regardless of placebo intake.

A two-page appendix on “patients excluded from the efficacy analysis” appeared 90 pages after the statistical analysis section ended, on page 2356. It was a table with no explanatory text listing 9 patients on placebo with their patient numbers and CDRSTL17 values.

CDRSTL17 values were likely CDRS-R scores at baseline, since the visit number corresponded to the randomisation visit for all 9 patients. But the table indicated that all 9 patients were excluded at the randomisation visit, which was true for only one of them. Scattered around in the report, information showed that 3 of the patients were discontinued after 7–10 days. The remaining 5 patients were not discontinued. According to the protocol and CSR, only 4 patients on placebo should have been excluded from the primary efficacy analysis, but another 5 were excluded, with no explanation why, while no fluoxetine patients were excluded. The difference between 0 and 9 exclusions is highly unlikely to have occurred by chance (P = 0.003, our calculation).

3.2.4. Washout phase

There was a two-week diagnostic evaluation period starting at week -3 (also called a “no drug” period in a figure), but it was not explained if previous antidepressant treatment was discontinued or if new drugs were not allowed. In a table of protocol violators 2405 pages later, a patient received the last dose of sertraline 12 days before the first visit, which should have been 14 days before. Only in this table could we see that it must have been a requirement to discontinue previous antidepressants 5 weeks before randomisation.

A one-week single blind placebo washout period started at week -1, but week -3 was the baseline for assessing the so-called placebo response at week 0. This did not make sense because placebo was not given before week -1.

Previous antidepressants were listed 6 times for fluoxetine and 6 for placebo, but as patients, drugs and adverse events were not linked, it was impossible to know if any of the serious adverse events in the introductory three-weeks period were iatrogenic withdrawal effects. Serious adverse events occurred in three patients who were not randomised: At week -3, one patient had explosive aggression, and another had psychotic symptoms; the third patient had suicidal ideation at week -1 coded as depression.

3.2.5. Concomitant medication

Judged by the clinical report forms, patients were not routinely asked which drugs they were currently on. After 447 pages of blank report forms, there was a form about concomitant medication investigators could use “at entry and during the study,” but as it was not obligatory to use it, information on other drugs used during the study was not reliable.

Concomitant medication was reported to have been used by more patients on fluoxetine than on placebo, 82% versus 66% during the first 9 weeks (P = 0.01), and 84% versus 72% during all 19 weeks (P = 0.03). Paracetamol was used more often (P = 0.02 and 0.04, respectively).

For drugs with sedative properties, the occurrences were: antihistamines 38 versus 31; sedatives/hypnotics 2 versus 5; antipsychotics 1 versus 0.

3.2.6. Effect of fluoxetine on depression

As in trial X065, all efficacy analyses were biased in favour of fluoxetine because the degree of depression after 9 and 19 weeks was unknown for discontinued patients. After two weeks, none had dropped out on fluoxetine versus 10 on placebo. Most analyses used the LOCF method, but Lilly did not alert its readers to the bias this caused.

FDA’s medical reviewer noted the considerably more dropouts on placebo than on fluoxetine and that the pattern was “rather unusual” because there were more dropouts on placebo than on fluoxetine for adverse events (9 versus 5), patient decision (11 versus 3) and lost to follow up (7 versus 1) [13]. Trial X065 had 0 dropouts for adverse events on placebo versus 5 on fluoxetine, and there were no losses to follow-up (see Table 3). FDA noted that Lilly had provided statements to the effect that their clinical investigators did not receive payment in return for particular results [13].

Many tables and graphs were confusing, with the type of analysis left unstated, and the terminology unclear, e.g. “from baseline to endpoint” could mean observed cases or LOCF.

The primary outcome result was described as: “A strong numerical trend was observed, with 71 (65%) fluoxetine-treated patients meeting response criteria compared with 54 (54%) placebo-treated patients; however, the difference between treatment groups was not statistically significant (P = 0.093) for this 9-week treatment period”. This analysis had no results for those 28% of the patients who were discontinued.

The difference in change scores on CDRS-R after 9 weeks was 5.9 in a graph whereas it was 7.2 in a table. In another graph, the difference was 7.2 using LOCF, which suggests that Lilly used LOCF also for the table without saying so.

Despite the flaws, placebo tended to be better than fluoxetine as evaluated by the patients: “Placebo-treated patients exhibited greater numerical reductions in the change from baseline for CDI and BDI total scores compared with fluoxetine-treated patients. The difference between treatment groups was not statistically significant”.

In the published article, the lack of an effect of fluoxetine on CDI and BDI was mentioned in these terms: “Given the high percentage of patients in this study who had comorbid ADHD, it may not be surprising that results of the clinician-rated measures were not reflected by the results of the patient-rated scales [8]”. However, only 14% of the children had an ADHD diagnosis, and it is not likely that such a diagnosis renders the children’s assessment of drug effects unreliable.

The report ended on page 224 but was succeeded by another 2325 pages with additional data. The psychiatrists used a CGI-Efficacy Index with 8 categories to “rate overall therapeutic effect in conjunction with side effects for each patient”. They assessed for each patient if the improvement in the depression outweighed any drug harms in terms of their interference with daily activities (which was not defined). Lilly claimed the results indicated that therapeutic effects outweighed any side effects because 58% versus 40% had a favourable score. We combined the data from the 8 categories by subtracting bad outcomes from good outcomes and found that 59% versus 55% had a good outcome (P = 0.58).

Fluoxetine was ineffective after 19 weeks for CDRS-R when observed cases were used (the difference to placebo was 2.5, P = 0.36). For the LOCF analysis, the difference was significant for all visits, even after the first week when patients on fluoxetine had only received 10 mg, but after 19 weeks, the difference was only 4.8 (P = 0.02). In an ANOVA of changes, the difference was 1.7 (P = 0.13).

In several cases, we could not understand how additional analyses differed from previous ones and why patient numbers differed when they should have been the same. In contrast to the 9-week results in the report summary (see below), several results were no longer statistically significant after 19 weeks, e.g. for MADRS, CGI-Improvement, response based on CGI-Improvement, recovery based on CDRS-R and CGI-Improvement scores, and CDRS-R mood subtotal. As an example, for CGI-Improvement, P = 0.03 after 9 weeks and P = 0.32 after 19 weeks.

There were narratives for 10% of the patients (6 in X065 and 24 in HCJE for post-randomisation events). As we explain below, these were very brief, unclear and incomplete. There were important errors, and two sets of tables separated by 668 pages needed to be combined.

The CSRs used two sets of terms interchangeably, the verbatim term reported by investigators and the coded term added by the company. Lilly did not specify which terms were used in three very different sets of tables. In one table, the column header was “Event Classification Term,” but what was reported was the verbatim term.

No patients died. Two patients had serious adverse events on fluoxetine and four on placebo, but the reporting was brief, opaque, and contradictory, and the terms were not the same across tables describing the same patients. Combining three tables and a figure it turned out that a patient on fluoxetine developed suicidal ideation 89 days before the first visit and was discontinued 70 days after randomisation when the suicidality had become serious, and the patient was hospitalised. The investigator considered the event possibly related to study drug, but Lilly coded it as depression. The suicidality became serious two weeks after some patients had been re-randomised to 40 mg of fluoxetine, but we could not assess its possible relation to the dose increase, as there was no information on dose. One of the tables indicated that this patient was withdrawn at randomisation and that fluoxetine could therefore not have caused the suicidality.

The other patient on fluoxetine had swollen tonsils coded as pharyngitis and underwent a tonsillectomy. The narrative revealed that the patient suffered from much else including fatigue and irritability.

The four serious adverse events on placebo were kidney infection, abdominal pain/appendicitis, aggressive behaviour coded as hostility, and self-mutilatory behaviour coded as intentional injury. The narrative for the patient with appendicitis used the term viral infection coded as infection, which was confusing, as these data were from the relapse prevention phase of the trial even though the company specified in numerous places that data from that phase would not be included in the CSR.

The patient with aggressive behaviour and the patient with self-mutilatory behaviour had symptoms at the first visit and were discontinued 10 and 37 days after randomisation, respectively, when the symptoms had become serious, and they were hospitalised. We had no information allowing us to judge if the events could be iatrogenic effects caused by withdrawal of an antidepressant before randomisation.

The information on the patient with aggressive behaviour appeared in a table, which mixed nonrandomised patients with patients with serious adverse events and patients withdrawn due to nonserious adverse events. The table header was nevertheless “All randomized patients”. As there were no dates, the patient may not have been randomised.

When self-mutilatory behaviour is coded as intentional injury, it conceals if it is violence towards self or others. A brief narrative of 65 words revealed that this patient had both suicidal ideation and homicidal ideation (coded as hostility) and was hospitalised for these reasons.

During the first 9 weeks, 19 versus 42 patients discontinued the study but Lilly’s report to the Data Monitoring Board two months after the study had been completed described 15 versus 33 discontinuations. These numbers were from week 7 without an explanation as to why the Board did not get the full results. This error was repeated for the 19 weeks results.

Eleven patients in each group discontinued due to adverse events, six of which were for nonserious psychiatric reasons on fluoxetine and one on placebo. The terms for fluoxetine were agitation (after 28 days), elevated mood coded as euphoria (65 days), physical aggression coded as hostility (98 days), hyperactivity coded as hyperkinesia (18 days), mania (32 days), and behavioural disinhibition coded as personality disorder (73 days).

According to the narratives, these patients were more severely affected than the tables suggested. The agitation was severe, and additional events included irritability, fatigue, decreased concentration, insomnia, anger, hearing voices, racing thoughts, mood swings and temper tantrums, which suggested drug induced psychosis and possibly akathisia.

The patient with euphoria was discontinued due to elevated mood, increased irritability, restlessness, increased hyperactivity and impulsivity, pressured speech, tangentiality, flight of ideas, belligerence, and mild looseness of associations. These events predispose to suicide and violence, which one would not suspect from the investigator term elevated mood or the coded term euphoria.

The patient with hyperactivity had an ADHD diagnosis. After 22 days on fluoxetine, the ADHD symptoms had become extreme, which may be drug induced akathisia.

The patient with mania also had irritability, agitation, insomnia, pressured speech, and delusions.

A seventh patient discontinued fluoxetine after 84 days due to endometrial hyperplasia. The 167-word narrative was only about this unusual diagnosis in a 15-year-old girl but the list of symptoms included akathisia.

An eighth patient discontinued fluoxetine after 120 days due to an increase in the severity of migraine. Bad dreams and aggression occurred 1-2 days before randomisation during the placebo washout. During treatment, sleep disturbance, shakiness, bad dreams, and intermittent restlessness coded as akathisia occurred.

A patient on placebo had anxiety at randomisation and discontinued 7 days later. Two days after randomisation this patient developed asthenia, and after seven days, he was “agitated and yelling at his mother. Patient stated that the study medication made him nervous, gave him a headache, and made him sick to his stomach. Patient demanded to be withdrawn from the study”. This patient could have suffered akathisia from withdrawal of prior medication.

In total, 7 fluoxetine versus 3 placebo patients experienced psychiatric adverse events leading to discontinuation, which become 9 versus 3 patients with significant psychiatric events if we add the two patients with akathisia on fluoxetine.

The information in a published paper from 2004, which was only about safety, disagreed markedly with the data in the CSR even though all five authors were from Lilly [14]. The paper stated that 49 versus 47 patients “completed 19 weeks of treatment” but the correct numbers were 40 versus 35. The paper noted that 4 patients on fluoxetine reported an event related to suicide or self-harm, but we found only one such patient in the CSR (who was hospitalized because of suicidal ideation). The three other patients, two with suicidal ideation and one with self-mutilation, were not described.

The paper also noted that 4 patients on placebo reported an event related to suicide or self-harm, but we found only one (who was hospitalized for suicidality and self-mutilation). One patient took 9 placebo capsules instead of 3, one made comments about wanting to die, and the third patient reported suicidal ideation.

These discrepancies suggest the company has access to information related to suicidality left out of the CSR, which stated that there was “no difference in the rate of reported suicide-related events between fluoxetine and placebo [14]”. This contradicted our findings (see Supplementary Table S1).

3.2.8. Other adverse events

Almost all patients experienced a solicited adverse event, 105 on fluoxetine versus 100 on placebo after 9 weeks (P = 0.17). The numbers after 19 weeks were 107 versus 101 (P = 0.06) in one analysis and 108 versus 102 (P = 0.04) in another.

For non-solicited adverse events, there were 94 on fluoxetine versus 80 on placebo after 9 weeks (P = 0.02) and 101 versus 87 after 19 weeks (P = 0.006).

After 9 weeks, more patients had experienced nervous system events on fluoxetine than on placebo, 35 versus 24 (P = 0.095). Lilly did not comment on this although, the P = 0.093 for the primary efficacy outcome was called a “strong numerical trend”. After 19 weeks, the difference was significant, 42 versus 28 patients (P = 0.01, number needed to harm 6). There were also more events related to the respiratory system, 61 versus 42 (P = 0.01) and to special senses, 16 versus 4 (P = 0.005).

There were no severity data in the main part of the CSR and Lilly did not define the three severity grades used. In drug trials, the usual definitions are:

Mild: awareness of sign or symptom, but easily tolerated.

For drugs that only have symptomatic effects, harms that interfere with usual activities are relevant for an assessment of the balance between benefits and harms. If only one comparison is chosen, it should therefore not be for severe events, as Lilly did, but for events classified as moderate or severe.

Lilly reported 19 versus 15 patients with one or more severe adverse events after 9 weeks (P = 0.46), whereas we found 68 versus 56 with moderate or severe events (P = 0.10). After 19 weeks, Lilly mentioned that there were not more patients with severe events on fluoxetine than on placebo, 22 versus 18, ignoring that 78 versus 64 patients had moderate or severe adverse events (P = 0.047, our calculation, number needed to harm 7).

Lilly downplayed even more the solicited adverse events remarking that “There were no clinically relevant differences in maximum intensity for any treatment-emergent solicited adverse event”. However, using Lilly’s terminology, there was a “strong” trend for severe adverse events after 9 weeks, 64 versus 52 (P = 0.105), which became 71 versus 57 after 19 weeks (P = 0.055). After 9 weeks, significantly more patients on fluoxetine than on placebo were feeling sleepy, 19 versus 7 (P = 0.01), having trouble getting along with parents, 19 versus 9 (P = 0.045), and trouble paying attention, 18 versus 7 (P = 0.02). Lilly stated that the differences were small and that the adverse events did not lead to discontinuations. However, the differences were about 10%, which means that for every 10 patients treated with fluoxetine, one was severely harmed. Nine versus 5 patients had severe problems with sitting still, which Lilly did not comment on although it could mean akathisia.

Fluoxetine reduced the increases in height and weight over 19 weeks by 1.0 cm and 1.1 kg, respectively (P = 0.008 for both). There were no data about these harms in the published article [8], and no comment on them in the CSR when they were presented [10].

The company concluded that “fluoxetine 20 to 60 mg/day is safe”. Discussing the 19-weeks data for all the patients, the CSR stated that the significance of the effect on height was uncertain, as previous studies had not found this. Lilly repeated this statement in their published paper on safety and added that the “Confidence in the interpretation of this result is limited, given that height in this study was not collected in a standardized fashion. Moreover, since measurements were recorded and rounded to the nearest inch, a small imprecision in measurement could potentially result in a 1-inch difference in recorded height [14]”. However, if recordings are rounded, this affects both groups. Furthermore, the concern is at odds with the fact that height for each patient was reported with two decimals (e.g. 160.02 cm) and weight with eight decimals (e.g. 47.62719885 kg) [14].

Fluoxetine increased the QTc interval by 6.95 msec (P = 0.02 for the difference to placebo). After numerous analyses and manipulations over six pages, e.g. showing that the P-value was not statistically significant for children with out-of-range ECG intervals and was not present in subgroup analyses by age, Lilly concluded that, “Two independent, blinded analyses of ECG interval changes did not reveal clinically significant changes in any ECG parameter”.

Finally, fluoxetine increased serum cholesterol (difference to placebo 0.2 mmol/L, P = 0.01).

3.2.9. Additional analyses of adverse events

In a “Safety Data Summary” of 1160 pages, for changes in height and weight, Lilly introduced a new interval, all randomised patients “who did not discontinue prior to Visit 11”. This visit was two weeks into the last 10-week period. There was no explanation why this interval was chosen. Data on ECGs were also presented using the new cut-off of 11 weeks. As before, the harmful effect of fluoxetine was removed by dichotomising the data and looking at out-of-range values.